1: \begin{abstract}

2: We revisit the classical dual ascent algorithm for minimization of convex functionals in the presence of linear constraints, and give convergence results which apply even for non-convex functionals. We describe limit points in terms of the convex envelope. We also introduce a new augmented version, which is shown to have superior convergence properties, and provide new results even for convex but non-differentiable objective functionals (as well as non-convex).

3: 

4: The results are applied to low rank approximation of a given matrix, subject to linear constraints. In particular, letting the linear constraints enforce Hankel structure of the respective matrices, the algorithms can be applied to complex frequency estimation. We provide numerical tests in this setting.

5: 

6: 

7: 

8: %For the problem of structured low rank approximation it is natural to consider the sum of the rank and the Frobenius error as an objective function. In the absence of a particular matrix structure of the approximation, this problem can be solved using singular value decomposition despite that it is non-convex. This is no longer true when a linear constraint is imposed on the matrix structure. A simple approach to address the structured low rank problem would be to use a standard Lagrangian multiplier to treat the structure constraint, and use the dual ascent method for an iterative approximate solution scheme. However, since the objective function is non-convex it does not seem at first sight that much can be said about the convergence of such a scheme. Nevertheless, we show in this paper that this scheme coincides with the scheme obtained by replacing the objective function with its convex envelope, and thus convergence results are derived. We illustrate the method on the problem of frequency estimation, where Hankel matrix structure is combined with low-rank approximation for the problem of approximating sampled function using sparse sums of exponentials.

9: \end{abstract}

10: